An ensemble ML model that predicts long-term relationship compatibility from two people's profiles, grounded in Gottman's Four Horsemen and Cox proportional hazards survival analysis.

👉 Try the live demo →

What this is (and isn't)

Is: A well-calibrated research artifact. An ensemble (XGBoost + LightGBM + CatBoost) trained on three open datasets, with Gottman behavioral proxies and survival priors layered in. Think of it as a mirror that reflects patterns the literature has documented — not a crystal ball.

Isn't: A decision tool. Don't break up, propose, or pick a partner based on its output. The interesting question isn't "what score did I get" — it's "which of the Four Horsemen showed up in my top factors, and why."

Training data is narrow: Columbia speed-daters (2002–2004), 170 Turkish couples from the Yöntem Gottman study, and 14,688 public-figure marriages pulled from a dataset originally compiled by Vedastro for unrelated research (we used only the marriage/divorce metadata — no astrological features). Generalization beyond these cohorts is unverified. See Limitations.

📊 Headline Results

Metric	v1.0 Baseline	v2.0 Enhanced	Change
AUC-ROC	0.8842	0.8896	+0.0055 ✅
AUC-PR	0.6933	0.7108	+0.0175 ✅
Brier Score	0.0960	0.0934	-0.0026 ✅
Accuracy	83.5%	85.9%	+2.4% ✅
F1 Score	0.620	0.630	+0.010 ✅
Precision	52.4%	58.9%	+6.5% ✅

Key improvement: +12.3% precision boost — far fewer false positives than v1.

What Changed

v2.0 adds 20 new features from two additional data sources:

Source	Features Added	Signal
Gottman Behavioral Model (Phase 1)	13	Contempt, criticism, defensiveness, stonewalling proxy scores derived from 170-couple divorce study
Marriage Duration Survival Model (Phase 2)	7	Longevity priors from 14,688 real marriages (age-risk, relationship-history risk, timing hazard)

8 of the 20 new features ranked in the top 30 most important features by SHAP:

Rank	New Feature	SHAP	Source
3	`gottman_proxy_love_maps`	0.447	🔴 Gottman
4	`gottman_proxy_contempt_x_stonewalling`	0.403	🔴 Gottman
8	`gottman_proxy_ratio`	0.306	🔴 Gottman
10	`gottman_proxy_stonewalling`	0.279	🔴 Gottman
12	`gottman_proxy_horsemen`	0.264	🔴 Gottman
21	`gottman_proxy_net_risk`	0.189	🔴 Gottman
27	`survival_age_gap_risk`	0.163	🔵 Survival
29	`gottman_proxy_contempt`	0.160	🔴 Gottman

Phase 1: Gottman Behavioral Model

Dataset: Yöntem et al. Divorce Predictors — 170 married/divorced Turkish couples, 54 Gottman-mapped behavioral questions.

Standalone performance: AUC = 0.998, Accuracy = 98.2% on predicting divorce from behavioral patterns.

The 54 questions map to Gottman's relationship theory:

Gottman Dimension	Questions	What It Measures
Shared Goals	Q1-Q10	Aligned life direction, quality time, common objectives
Love Maps	Q11-Q20	Values alignment, role expectations, compatibility beliefs
Love Maps Deep	Q21-Q30	Knowing partner's inner world, stress, hopes, anxieties
Criticism	Q31-Q32, Q37-Q38	Attacking character, negative statements, sudden arguments
Contempt	Q33-Q36, Q39-Q40	Insults, humiliation, anger escalation, hatred
Defensiveness	Q41, Q45-Q46, Q48-Q50	Blame-shifting, victimhood, refusing responsibility
Stonewalling	Q42-Q44, Q47	Silence, withdrawal, leaving, shutting down
Deep Contempt	Q51-Q54	Attributing meanness, vindictiveness, pathology to partner

Top divorce predictor by SHAP: love_maps × shared_goals interaction — couples who both lack shared goals and don't know each other's inner world face the highest divorce risk.

Gottman Proxy Features (mapped to speed dating data)

Since speed dating participants didn't answer the 54 Gottman questions, we created proxy scores by mapping their existing personality/perception data to Gottman dimensions:

Proxy	Derived From
`gottman_proxy_contempt`	Low mutual scores + high perception gaps
`gottman_proxy_criticism`	Misaligned values + asymmetric ratings
`gottman_proxy_defensiveness`	Self-rating inflation vs partner perception
`gottman_proxy_stonewalling`	Low engagement, low liking, no shared interests
`gottman_proxy_love_maps`	Interest correlation + shared interests + mutual perception accuracy
`gottman_proxy_shared_goals`	Value alignment + interest overlap
`gottman_proxy_ratio`	The famous Gottman 5:1 positive-to-negative ratio

Phase 2: Marriage Duration Survival Model

Dataset: vedastro-org/15000-Famous-People-Marriage-Divorce-Info — 14,688 marriage records from 12,353 famous people.

Key Findings

Finding	Statistic
Overall divorce rate	34.5%
Median divorce timing	7 years
Most dangerous period	3-7 years (41.1% of all divorces)
Love marriage divorce rate	34.1%
Arranged marriage divorce rate	23.4% (p=0.006, significantly lower)
First marriage divorce rate	27.8%
Subsequent marriage divorce rate	69.3%

Cox Proportional Hazards Model (Concordance = 0.64)

Factor	Hazard Ratio	p-value	Meaning
Is first marriage	0.26	<0.001	74% lower divorce hazard than subsequent marriages
Is love marriage	0.77	0.002	23% lower hazard than non-love marriages
Age at marriage	0.96	<0.001	Each year older → 4% lower divorce hazard
Marriage number	1.34	<0.001	Each additional marriage → 34% higher hazard

Divorce Timing Distribution

Kaplan-Meier Survival Curves

Model Architecture (v2.0)

Ensemble of 3 gradient-boosted tree models with 133 engineered features (113 original + 13 Gottman + 7 survival):

Model	Weight	v1 AUC	v2 AUC	Change
XGBoost	0.40	0.8852	0.8920	+0.0068
LightGBM	0.35	0.8912	0.9011	+0.0099
CatBoost	0.25	0.8661	0.8688	+0.0027
Ensemble	—	0.8842	0.8896	+0.0055

Visualizations

v1 vs v2 ROC Comparison

Metrics Comparison

Feature Source Contribution

Enhanced SHAP Summary (v2)

v1 Visualizations


ROC Curves	SHAP Summary
Feature Importance	Confusion Matrix

Training Data

Dataset	Records	Role
mstz/speeddating	1,048 encounters	Primary training data — individual profiles + match outcome
Yöntem et al. Divorce Predictors (Kaggle)	170 couples	Phase 1 — Gottman behavioral feature engineering
vedastro-org/15000-Famous-People-Marriage-Divorce-Info	14,688 marriages	Phase 2 — Longevity priors + survival analysis

Literature Basis

Paper	Contribution
Grinsztajn et al. (NeurIPS 2022) — "Why do tree-based models still outperform deep learning on tabular data?"	Validated XGBoost/LightGBM as SOTA for medium-sized tabular data
Fisman et al. (QJE 2006) — "Gender Differences in Mate Selection"	Original speed dating experiment; ~70% accuracy with logistic regression
Gottman & Silver (1999) — "The Seven Principles for Making Marriage Work"	Four Horsemen framework: contempt, criticism, defensiveness, stonewalling
Yöntem et al. (2019) — "Divorce Prediction Using Correlation Based Feature Selection"	54-question Gottman-mapped divorce predictor; published 97.7% accuracy
Savcisens et al. (Nature Human Behaviour 2024) — "Using Sequences of Life-events to Predict Human Lives"	life2vec — longitudinal prediction architecture

Repo Structure

├── # v1.0 Baseline Model
├── xgboost_model.joblib, lightgbm_model.joblib, catboost_model.cbm
├── ensemble_config.json, feature_columns.joblib
├── figures/                          # v1 plots
│
├── # Phase 1 — Gottman Behavioral Model
├── phase1_divorce_model/
│   ├── divorce_xgb.joblib, divorce_lgb.joblib, divorce_cat.cbm
│   ├── gottman_recipe.json           # Dimension mappings + importance
│   ├── gottman_mapping.joblib
│   └── figures/                      # SHAP, confusion matrix, dimension importance
│
├── # Phase 2 — Survival Model
├── phase2_survival_model/
│   ├── longevity_priors.json         # Base rates by type/era/age/marriage#
│   ├── survival_recipe.json          # Cox PH + KM + timing distributions
│   └── figures/                      # KM curves, Cox hazard ratios, timing
│
├── # v2.0 Enhanced Model (RECOMMENDED)
├── v2_enhanced/
│   ├── enhanced_xgb.joblib, enhanced_lgb.joblib, enhanced_cat.cbm
│   ├── enhanced_config.json          # Weights, features, metrics, improvements
│   ├── enhanced_feature_columns.joblib
│   └── figures/                      # Comparison plots, SHAP
│
└── # Training Scripts (fully reproducible)
    ├── train_relationship_predictor.py   # v1 baseline
    ├── phase1_divorce_model.py           # Gottman behavioral model
    ├── phase2_marriage_duration.py       # Survival analysis
    └── phase3_integration.py             # Integration + comparison

Limitations & Ethics

Cohort bias. The primary training signal is from Columbia University speed-daters in 2002–2004. This is a narrow demographic slice — predominantly educated, urban, US-based, early-internet-era. Generalization to other populations is unverified and should be assumed weak until tested.

Celebrity bias in the survival priors. The 14,688-marriage Vedastro dataset is public-figure-heavy, with known elevated divorce rates and atypical relationship dynamics (media exposure, wealth asymmetry, career mobility). The arranged-vs-love finding (23.4% vs 34.1%) is descriptive of this dataset, not a general claim about relationship types.

Dataset provenance. The Vedastro dataset was originally compiled for astrology research. This model uses only the structured marriage/divorce metadata (age at marriage, marriage number, duration, type, outcome) — no astrological variables are used as features.

Short-horizon proxy. Speed-dating captures initial match decisions, not long-term outcomes. The Gottman and survival layers partially bridge this gap, but they're proxies, not ground truth.

Small Gottman sample. The underlying divorce predictor was trained on 170 couples. The Four Horsemen framework itself is robust across decades of research; the proxy mapping from speed-dating features to Gottman dimensions is approximate and worth questioning.

Not a decision tool. Outputs are probabilistic, directional, and should be treated as a conversation starter — not advice. This model should not be used to make real decisions about real relationships.

License

cc-by-nc-4.0 Research use. Based on publicly available academic datasets.

Built with XGBoost, LightGBM, CatBoost, SHAP, lifelines, and scikit-learn.

Builder-Neekhil/relationship-longevity-predictor

💕 Relationship Longevity Predictor — v2.0